Add support for RGBA128F/ RGB96F
I've just tested this and floating point format seems to work as intended.
This requires pixman 0.36 to enable support, but it looks like it's not enough yet to blend correctly. Pixman 0.36 still falls back to expanding 8-bit values, effectively limiting the extra resolution. However, while I work on that, the floating point support conversion actually works, so I wanted to submit this to cairo while I fix pixman in parallel.